Have you ever had a moment where a word was on the tip of your tongue, but you couldn’t quite remember it? For a bit of fun, here are two questions that might generate this feeling for you:
- What is the traditional, long, wooden, Aboriginal wind instrument most associated with the Indigenous peoples in Australia?
2. What is the name for mammals like kangaroos that carry their young in their pouch?
Before you open a new tab to search for the answer if it doesn’t come to mind, take a moment to sit with this tip-of-the-tongue feeling. Can you think of the first letter? The number of syllables? A word that rhymes with it?
As a psycholinguist myself, I love examples like this that probe at the mind’s capacity for language. This specific psycholinguistic phenomenon, called the tip-of-the-tongue state, is when some information about a word can be recalled without remembering the full word. It’s an interesting sensation because it teaches us a lot about the mental lexicon, the store of words, and their associated meanings in our minds.
Many people might think of the mental lexicon as a mental dictionary: a list of words and their definitions. In this metaphor, when you try to remember a word, you consult your list and search for the word you want. Children learning words can simply record a new entry and slot it into the list! But the tip-of-the-tongue state shows that we can recall partial information about a word, such as meaning, syllable length, or starting letter, without recalling the word itself. (For those who followed along and still can’t find the answers to the questions above, the instrument is a didgeridoo, and the animal category is marsupial.)
Or, have you ever mixed up two similar-sounding words and said something like “the pineapple of success” or “photogenic memory”? Even though “pineapple” and “pinnacle” share no aspect of meaning, sharing sounds is enough for us to mix up our words, so words in our mental lexicon must also be connected by sounds as well as meaning. Taking this together with tip-of-the-tongue states, word information must be distributed throughout the mental lexicon rather than occur as a static dictionary entry. But what would that look like?
In a recent publication in Psychonomic Bulletin and Review, Massimo Stella, Salvatore Citraro, Giulio Rossetti, Daniele Marinazzo, Yoed N. Kenett, and Michael S. Vitevitch (pictured below) discuss an alternate way of modeling the mental lexicon: as a multilayer network.
Multilayer networks use layers of nodes and links between nodes or layers to represent different pieces of information about a word. For example, the word “cat” might involve nodes for each of the sounds (‘k,’ ‘a’, and ‘t’) in one layer, then connect to form “cat” in another layer, which connects to nodes in another layer for “animal” and “likes to knock my cup off the table.” The whole concept of “cat” is represented as the pattern of activation across all these layers. In this example, I used three layers, but other models might add layers for other linguistic properties like the word’s syntactic properties.
Instead of distributing properties across the various layers, a multiplex network model emphasizes the connections between different words on different layers by keeping the same nodes in every layer. For example, “cat” would connect to “mat” and “pat” at the sound layer since they share phonemes, connect to “kitty” and “feline” at the semantic layer since they share meanings, and connect to “dog” and “scaredy” at the association layer. Imagine a spiderweb of interconnected words where connecting strands depend on the layer of the network. The phoneme web connects similar sounding words while the meaning web connects words with different meanings. Just like a multilayer network, a multiplex network can have much more than the three layers in our example, but Stella and the authors note that the number and type of layers should be motivated by a balance between maximizing accuracy and minimizing redundancy.
The authors emphasize that these albeit complex models perform much better at explaining real-world data than other models. Remember tip-of-the-tongue states? Recalling partial information about a concept can be modeled in multilayer or multiplex models by activating the nodes and connections for the parts of the concept accounting for initial sounds and meaning but not the full word.
As opposed to “black box” artificial neural networks, multiplex models are more easily interpretable by researchers and explain other psycholinguistic findings. For one, “communities” of words arise when groups of words in the network are closely linked to each other. Accounting for how long, complex, or rare words are, the multiplex model shows that children regularly acquire one special community of words all at once around seven or eight years old, while words outside the community are only learned gradually over time. The authors also call out this same special community as explaining which words are more likely to be retained in patients with aphasia or more likely to be accessed by individuals with different levels of creativity searching through their mental lexicon.
The authors compare multilayer networks to the physics of light refraction,
“like prisms decompose white light in multiple colours, cognitive multilayer networks can highlight multiple, distinct associations between concepts.”
My takeaway: language is beautiful and complex, and our models should reflect that.
Featured Psychonomic Society Article
Stella, M., Citraro, S., Rossetti, G., Marinazzo, D., Kenett, Y. N., & Vitevitch, M. S. (2024). Cognitive modelling of concepts in the mental lexicon with multilayer networks: Insights, advancements, and future challenges. Psychonomic Bulletin & Review, 1-24. https://doi.org/10.3758/s13423-024-02473-9
